[Dist][EP] Remove ETP/EP maintained in vllm-ascend (#1681)

### What this PR does / why we need it? Remove ETP/EP maintained in branch main. We drop this as there is no relevant scenarios to use ETP now, and we may subsequently advocate implementing expert tensor parallelism in vLLM to support scenarios where the expert is needed to be sliced This is a part of #1422 backport. Fixes https://github.com/vllm-project/vllm-ascend/issues/1396 https://github.com/vllm-project/vllm-ascend/issues/1154 ### Does this PR introduce _any_ user-facing change? We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in vllm instead. ### How was this patch tested? CI passed with new added and existing test. - vLLM version: v0.9.2 - vLLM main: fe8a2c544a Signed-off-by: MengqingCao <cmq0113@163.com>
2025-07-21 09:08:04 +08:00
parent a8b316ac5b
commit 8cfd257992
24 changed files with 66 additions and 548 deletions
--- a/tests/e2e/long_term/accuracy/accuracy_multicard.py
+++ b/tests/e2e/long_term/accuracy/accuracy_multicard.py
@@ -36,7 +36,7 @@ COMPLETIONS_URL = f"http://{SERVER_HOST}:{SERVER_PORT}/v1/completions"

 # pre-trained model path on Hugging Face.
 # Qwen/Qwen2.5-0.5B-Instruct: accuracy test for DP.
-# Qwen/Qwen3-30B-A3B: accuracy test for EP and ETP.
+# Qwen/Qwen3-30B-A3B: accuracy test for EP.
 # deepseek-ai/DeepSeek-V2-Lite: accuracy test for TP.
 MODEL_NAME = ["Qwen/Qwen3-30B-A3B", "deepseek-ai/DeepSeek-V2-Lite"]

@@ -200,62 +200,3 @@ def test_lm_eval_accuracy_dp(model, max_tokens):
        except subprocess.TimeoutExpired:
            server_proc.kill()
            server_proc.wait()
-
-
-@pytest.mark.parametrize("max_tokens", [10])
-@pytest.mark.parametrize("model", ["Qwen/Qwen3-30B-A3B"])
-def test_lm_eval_accuracy_etp(model, max_tokens):
-    log_file = open("accuracy_etp.log", "a+")
-    cmd = [
-        "vllm", "serve", model, "--max_model_len", "4096",
-        "--tensor_parallel_size", "4", "--enforce_eager",
-        "--enable_expert_parallel", "--additional_config",
-        '{"expert_tensor_parallel_size": "4"}'
-    ]
-    server_proc = subprocess.Popen(cmd,
-                                   stdout=log_file,
-                                   stderr=subprocess.DEVNULL)
-
-    try:
-        for _ in range(300):
-            try:
-                r = requests.get(HEALTH_URL, timeout=1)
-                if r.status_code == 200:
-                    break
-            except requests.exceptions.RequestException:
-                pass
-            time.sleep(1)
-        else:
-            log_file.flush()
-            log_file.seek(0)
-            log_content = log_file.read()
-            pytest.fail(
-                f"vLLM serve did not become healthy after 300s: {HEALTH_URL}\n"
-                f"==== vLLM Serve Log Start ===\n{log_content}\n==== vLLM Serve Log End ==="
-            )
-
-        prompt = "bejing is a"
-        payload = {
-            "prompt": prompt,
-            "max_tokens": max_tokens,
-            "sampling_params": {
-                "temperature": 0.0,
-                "top_p": 1.0,
-                "seed": 123
-            }
-        }
-        resp = requests.post(COMPLETIONS_URL, json=payload, timeout=30)
-        resp.raise_for_status()
-        data = resp.json()
-
-        generated = data["choices"][0]["text"].strip()
-        expected = "city in china. it is the capital city of"
-        assert generated == expected, f"Expected `{expected}`, got `{generated}`"
-
-    finally:
-        server_proc.send_signal(signal.SIGINT)
-        try:
-            server_proc.wait(timeout=10)
-        except subprocess.TimeoutExpired:
-            server_proc.kill()
-            server_proc.wait()
--- a/tests/e2e/multicard/test_expert_parallel.py
+++ b/tests/e2e/multicard/test_expert_parallel.py
@@ -0,0 +1,30 @@
+import pytest
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.model_utils import check_outputs_equal
+
+
+@pytest.mark.parametrize("model_name", ["deepseek-ai/DeepSeek-V2-Lite-Chat"])
+def test_e2e_ep_correctness(model_name):
+    example_prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    max_tokens = 5
+
+    with VllmRunner(model_name, tensor_parallel_size=2) as vllm_model:
+        tp_output = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    with VllmRunner(model_name,
+                    tensor_parallel_size=2,
+                    enable_expert_parallel=True) as vllm_model:
+        ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=ep_output,
+        outputs_1_lst=tp_output,
+        name_0="ep_output",
+        name_1="tp_output",
+    )
--- a/tests/e2e/multicard/test_fused_moe_allgather_ep.py
+++ b/tests/e2e/multicard/test_fused_moe_allgather_ep.py
@@ -50,7 +50,6 @@ def test_generate_with_allgather():
                            "enabled": True,
                            "chunked_prefill_enabled": False,
                        },
-                        "expert_tensor_parallel_size": 1
                    }) as vllm_model:
        vllm_model.generate(example_prompts, sampling_params)

@@ -74,6 +73,5 @@ def test_generate_with_alltoall():
                            "enabled": True,
                            "chunked_prefill_enabled": False,
                        },
-                        "expert_tensor_parallel_size": 1
                    }) as vllm_model:
        vllm_model.generate(example_prompts, sampling_params)
--- a/tests/e2e/multicard/test_torchair_graph_mode.py
+++ b/tests/e2e/multicard/test_torchair_graph_mode.py
@@ -123,6 +123,7 @@ def _pangu_torchair_test_fixture(
            distributed_executor_backend="mp",
            enforce_eager=False,
            additional_config=additional_config,
+            enable_expert_parallel=True,
    ) as vllm_model:
        # use greedy sampler to make sure the generated results are fix
        vllm_output = vllm_model.generate_greedy(example_prompts, 5)
--- a/tests/ut/distributed/test_parallel_state.py
+++ b/tests/ut/distributed/test_parallel_state.py
@@ -1,208 +0,0 @@
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# This file is a part of the vllm-ascend project.
-#
-
-from unittest.mock import MagicMock, patch
-
-import pytest
-from vllm.distributed.parallel_state import GroupCoordinator
-
-import vllm_ascend
-from tests.ut.base import TestBase
-from vllm_ascend.distributed.parallel_state import (
-    destory_ascend_model_parallel, get_ep_group, get_etp_group,
-    init_ascend_model_parallel, model_parallel_initialized)
-
-
-class TestParallelState(TestBase):
-
-    @patch('vllm_ascend.distributed.parallel_state._EP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    def test_get_ep_group_when_initialized(self, mock_ep):
-        # Act
-        result = get_ep_group()
-
-        # Assert
-        assert isinstance(result, GroupCoordinator)
-
-    @patch('vllm_ascend.distributed.parallel_state._EP', None)
-    def test_get_ep_group_when_not_initialized(self):
-        # Act & Assert
-        with pytest.raises(AssertionError) as excinfo:
-            get_ep_group()
-        assert "expert model parallel group is not initialized" in str(
-            excinfo.value)
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    def test_get_etp_group_when_initialized(self, mock_etp):
-        # Act
-        result = get_etp_group()
-
-        # Assert
-        assert isinstance(result, GroupCoordinator)
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP', None)
-    def test_get_etp_group_when_not_initialized(self):
-        # Act & Assert
-        with pytest.raises(AssertionError) as excinfo:
-            get_etp_group()
-        assert "expert tensor parallel group is not initialized" in str(
-            excinfo.value)
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP', None)
-    @patch('vllm_ascend.distributed.parallel_state._EP', None)
-    def test_model_parallel_initialized_when_both_none(self):
-        # Act & Assert
-        assert not model_parallel_initialized()
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    @patch('vllm_ascend.distributed.parallel_state._EP', None)
-    def test_model_parallel_initialized_when_ep_none(self, mock_etp):
-        # Act & Assert
-        assert not model_parallel_initialized()
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP', None)
-    @patch('vllm_ascend.distributed.parallel_state._EP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    def test_model_parallel_initialized_when_etp_none(self, mock_ep):
-        # Act & Assert
-        assert not model_parallel_initialized()
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    @patch('vllm_ascend.distributed.parallel_state._EP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    def test_model_parallel_initialized_when_etp_initialized(
-            self, mock_ep, mock_etp):
-        # Act & Assert
-        assert model_parallel_initialized()
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    @patch('vllm_ascend.distributed.parallel_state._EP',
-           new_callable=lambda: MagicMock(spec=GroupCoordinator))
-    def test_destroy_when_both_exist(self, mock_ep, mock_etp):
-        # Act
-        destory_ascend_model_parallel()
-        # Assert
-        mock_ep.destroy.assert_called_once()
-        mock_etp.destroy.assert_called_once()
-        assert vllm_ascend.distributed.parallel_state._ETP is None
-        assert vllm_ascend.distributed.parallel_state._EP is None
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP', None)
-    @patch('vllm_ascend.distributed.parallel_state._EP',
-           new_callable=lambda: MagicMock())
-    def test_destory_ascend_model_parallel_when_etp_none(self, mock_ep):
-        # Act
-        destory_ascend_model_parallel()
-        # Assert
-        mock_ep.destroy.assert_called_once()
-        assert vllm_ascend.distributed.parallel_state._EP is None
-        assert vllm_ascend.distributed.parallel_state._ETP is None
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP',
-           new_callable=lambda: MagicMock())
-    @patch('vllm_ascend.distributed.parallel_state._EP', None)
-    def test_destory_ascend_model_parallel_when_ep_none(self, mock_etp):
-        # Act
-        destory_ascend_model_parallel()
-        # Assert
-        mock_etp.destroy.assert_called_once()
-        assert vllm_ascend.distributed.parallel_state._ETP is None
-        assert vllm_ascend.distributed.parallel_state._EP is None
-
-    @patch('vllm_ascend.distributed.parallel_state._ETP', None)
-    @patch('vllm_ascend.distributed.parallel_state._EP', None)
-    def test_destory_ascend_model_parallel_when_both_none(self):
-        # Act
-        destory_ascend_model_parallel()
-        # Assert
-        assert vllm_ascend.distributed.parallel_state._ETP is None
-        assert vllm_ascend.distributed.parallel_state._EP is None
-
-    @patch('torch.distributed.is_initialized', return_value=True)
-    @patch('torch.distributed.get_world_size', return_value=8)
-    @patch('vllm_ascend.distributed.parallel_state.get_world_group',
-           return_value=MagicMock(device_group='npu:0', local_rank=0))
-    @patch('torch.distributed.get_backend', return_value='hccl')
-    @patch('vllm_ascend.distributed.parallel_state.init_model_parallel_group')
-    @patch('vllm_ascend.distributed.parallel_state.model_parallel_initialized',
-           return_value=False)
-    def test_init_ascend_model_parallel_normal_case(
-            self, mock_mp_init, mock_init_group, mock_get_backend,
-            mock_world_group, mock_get_world_size, mock_is_init):
-        """Test normal initialization with default parameters"""
-        # Act
-        init_ascend_model_parallel()
-        # Assert
-        mock_init_group.assert_any_call([[0, 1, 2, 3, 4, 5, 6, 7]],
-                                        0,
-                                        'hccl',
-                                        group_name="ep")
-        mock_init_group.assert_any_call([[0]], 0, 'hccl', group_name="etp")
-        self.assertIsNotNone(vllm_ascend.distributed.parallel_state._EP)
-        self.assertIsNotNone(vllm_ascend.distributed.parallel_state._ETP)
-
-    @patch('vllm_ascend.distributed.parallel_state.model_parallel_initialized',
-           return_value=True)
-    def test_init_ascend_model_parallel_skip_if_initialized(
-            self, mock_mp_init):
-        """Test skipping when model parallel already initialized"""
-        with patch.object(vllm_ascend.distributed.parallel_state,
-                          '_EP') as mock_ep, patch.object(
-                              vllm_ascend.distributed.parallel_state,
-                              '_ETP') as mock_etp:
-            # Act
-            init_ascend_model_parallel()
-            # Assert
-            mock_ep.assert_not_called()
-            mock_etp.assert_not_called()
-
-    @patch('torch.distributed.is_initialized', return_value=False)
-    def test_init_ascend_model_parallel_assert_dist_not_init(
-            self, mock_is_init):
-        """Test assertion when distributed not initialized"""
-        # Act & Assert
-        with self.assertRaises(AssertionError):
-            init_ascend_model_parallel()
-
-    @patch('torch.distributed.is_initialized', return_value=True)
-    @patch('torch.distributed.get_world_size', return_value=8)
-    @patch('vllm_ascend.distributed.parallel_state.get_world_group',
-           return_value=MagicMock(device_group='npu:0', local_rank=1))
-    @patch('torch.distributed.get_backend', return_value='hccl')
-    @patch('vllm_ascend.distributed.parallel_state.init_model_parallel_group')
-    @patch('vllm_ascend.distributed.parallel_state.model_parallel_initialized',
-           return_value=False)
-    def test_init_ascend_model_parallel_custom_params(
-            self, mock_mp_init, mock_init_group, mock_get_backend,
-            mock_world_group, mock_get_world_size, mock_is_init):
-        """Test initialization with custom parallel sizes"""
-        # Act
-        init_ascend_model_parallel(expert_parallel_size=2,
-                                   expert_tensor_parallel_size=4,
-                                   world_size=8,
-                                   backend='hccl')
-        #Assert
-        mock_init_group.assert_any_call([[0, 4], [1, 5], [2, 6], [3, 7]],
-                                        1,
-                                        'hccl',
-                                        group_name="ep")
-        mock_init_group.assert_any_call([[0, 1, 2, 3], [4, 5, 6, 7]],
-                                        1,
-                                        'hccl',
-                                        group_name="etp")
--- a/tests/ut/test_ascend_config.py
+++ b/tests/ut/test_ascend_config.py
@@ -42,7 +42,6 @@ class TestAscendConfig(TestBase):
        test_vllm_config = VllmConfig()
        # No additional config given, check the default value here.
        ascend_config = init_ascend_config(test_vllm_config)
-        self.assertEqual(ascend_config.expert_tensor_parallel_size, 0)
        self.assertIsNone(ascend_config.expert_map_path)

        torchair_graph_config = ascend_config.torchair_graph_config
@@ -75,12 +74,10 @@ class TestAscendConfig(TestBase):
            "ascend_scheduler_config": {
                "enabled": True
            },
-            "expert_tensor_parallel_size": 1,
            "expert_map_path": "test_expert_map_path",
            "refresh": True
        }
        ascend_config = init_ascend_config(test_vllm_config)
-        self.assertEqual(ascend_config.expert_tensor_parallel_size, 1)
        self.assertEqual(ascend_config.expert_map_path, "test_expert_map_path")

        torchair_graph_config = ascend_config.torchair_graph_config
--- a/tests/ut/test_platform.py
+++ b/tests/ut/test_platform.py
@@ -28,7 +28,6 @@ class TestNPUPlatform(TestBase):
        self.mock_vllm_config.speculative_config = None

        self.mock_ascend_config = MagicMock()
-        self.mock_ascend_config.expert_tensor_parallel_size = 0
        self.mock_ascend_config.torchair_graph_config.enabled = False
        self.mock_ascend_config.ascend_scheduler_config.enabled = False

@@ -253,30 +252,6 @@ class TestNPUPlatform(TestBase):
        mock_init_ascend.assert_called_once_with(self.mock_vllm_config)
        mock_check_ascend.assert_called_once()

-    @patch("vllm_ascend.utils.is_310p", return_value=False)
-    @patch("vllm_ascend.ascend_config.check_ascend_config")
-    @patch("vllm_ascend.ascend_config.init_ascend_config")
-    def test_check_and_update_config_expert_parallel_enabled(
-            self, mock_init_ascend, mock_check_ascend, mock_is_310p):
-        mock_init_ascend.return_value = self.mock_ascend_config
-        self.mock_vllm_config.parallel_config.enable_expert_parallel = True
-        self.mock_vllm_config.parallel_config.tensor_parallel_size = 2
-        self.mock_vllm_config.parallel_config.world_size_across_dp = 4
-
-        from vllm_ascend import platform
-
-        importlib.reload(platform)
-
-        self.platform.check_and_update_config(self.mock_vllm_config)
-
-        self.assertEqual(
-            self.mock_vllm_config.parallel_config.expert_tensor_parallel_size,
-            1)
-        self.assertEqual(
-            self.mock_vllm_config.parallel_config.expert_parallel_size,
-            self.mock_vllm_config.parallel_config.world_size_across_dp,
-        )
-
    @patch("vllm_ascend.utils.is_310p", return_value=False)
    @patch("vllm_ascend.ascend_config.check_ascend_config")
    @patch("vllm_ascend.ascend_config.init_ascend_config")